tensorwatch/saliency/inverter_util.py [296:328]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
            new_norm[:, :out_c] = norm[:, :out_c] + norm[:, out_c:2 * out_c]
            new_norm[:, out_c:] = norm[:, 2 * out_c:3 * out_c] + norm[:, 3 * out_c:]
            norm = new_norm

            # Some 'rare' neurons only receive either
            # only positive or only negative inputs.
            # Conservation of relevance does not hold, if we also
            # rescale those neurons by (1+beta) or -beta.
            # Therefore, catch those first and scale norm by
            # the according value, such that it cancels in the fraction.

            # First, however, avoid NaNs.
            mask = norm == 0
            # Set the norm to anything non-zero, e.g. 1.
            # The actual inputs are zero at this point anyways, that
            # is why norm is zero in the first place.
            norm[mask] = 1
            # The norm in the b-rule has shape (N, 2*out_c, *spatial_dims).
            # The first out_c block corresponds to the positive norms,
            # the second out_c block corresponds to the negative norms.
            # We find the rare neurons by choosing those nodes per channel
            # in which either the positive norm ([:, :out_c]) is zero, or
            # the negative norm ([:, :out_c]) is zero.
            rare_neurons = (mask[:, :out_c] + mask[:, out_c:])

            # Also, catch new possibilities for norm == zero to avoid NaN..
            # The actual value of norm again does not really matter, since
            # the pre-factor will be zero in this case.

            norm[:, :out_c][rare_neurons] *= 1 if self.beta == -1 else 1 + self.beta
            norm[:, out_c:][rare_neurons] *= 1 if self.beta == 0 else -self.beta
            # Add stabilizer term to norm to avoid numerical instabilities.
            norm += self.eps * torch.sign(norm)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



tensorwatch/saliency/inverter_util.py [454:485]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                new_norm[:, :out_c] = norm[:, :out_c] + norm[:, out_c:2 * out_c]
                new_norm[:, out_c:] = norm[:, 2 * out_c:3 * out_c] + norm[:, 3 * out_c:]
                norm = new_norm
                # Some 'rare' neurons only receive either
                # only positive or only negative inputs.
                # Conservation of relevance does not hold, if we also
                # rescale those neurons by (1+beta) or -beta.
                # Therefore, catch those first and scale norm by
                # the according value, such that it cancels in the fraction.

                # First, however, avoid NaNs.
                mask = norm == 0
                # Set the norm to anything non-zero, e.g. 1.
                # The actual inputs are zero at this point anyways, that
                # is why norm is zero in the first place.
                norm[mask] = 1
                # The norm in the b-rule has shape (N, 2*out_c, *spatial_dims).
                # The first out_c block corresponds to the positive norms,
                # the second out_c block corresponds to the negative norms.
                # We find the rare neurons by choosing those nodes per channel
                # in which either the positive norm ([:, :out_c]) is zero, or
                # the negative norm ([:, :out_c]) is zero.
                rare_neurons = (mask[:, :out_c] + mask[:, out_c:])

                # Also, catch new possibilities for norm == zero to avoid NaN..
                # The actual value of norm again does not really matter, since
                # the pre-factor will be zero in this case.

                norm[:, :out_c][rare_neurons] *= 1 if self.beta == -1 else 1 + self.beta
                norm[:, out_c:][rare_neurons] *= 1 if self.beta == 0 else -self.beta
                # Add stabilizer term to norm to avoid numerical instabilities.
                norm += self.eps * torch.sign(norm)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -